Self-Steering Language Models
Gabriel Grand, Joshua B. Tenenbaum, Vikash K. Mansinghka, Alexander K. Lew, Jacob Andreas·April 09, 2025
Summary
DISCIPL, a self-steering language model method, outperforms large models like GPT-4 on complex tasks without fine-tuning. It combines a Planner with Follower models, enhancing efficiency, controllability, and commonsense reasoning. Evaluated on the PUZZLES dataset, DISCIPL excels in metrics like Pass@1, coherency, and error rates. The system uses a BaseModel for text generation with constraints, employing a hint method for updating proposals. It features CollieModel classes for generating sentences under specific length and word constraints. One model creates an 82-character sentence, while another produces a 11-word sentence with fixed words at specific positions. Both models use hints for remaining length and enforce token limits. A class generates a 3-day Singapore itinerary, incorporating task-specific variables, time range extraction, and a step method for itinerary creation, ensuring each day includes a time range and activity, with conditions for ending the itinerary.
Advanced features